Workflows/Create Uncensored AI Bots with Rakazo

Agents / WORKFLOW

Create Uncensored AI Bots with Rakazo

Use OrcaSAQ-2-Cyber-27B-Uncensored-GGUF with Rakazo locally through Ollama or remotely through a GGUF-enabled RunPod Serverless vLLM worker.

Verified sourceSource checked 10/17/2026

LOCALRENTED GPU

Before you begin

Hardware

Both routes use the exact 15.7 GB OrcaSAQ GGUF. Local and vLLM memory overhead differ. Start at 8192 context with one bot; these are estimates, not measured performance.

Machine	Local Ollama	RunPod Serverless vLLM
16 GB GPU + 32 GB system RAM	Tight; CPU offload may slow replies	Local GPU unused
24–32 GB GPU + 32–48 GB system RAM	Recommended local starting point	Local GPU unused
Apple Silicon · 24 GB unified memory	Tight with Docker and macOS	Local model memory unnecessary
Apple Silicon · 32–64 GB unified memory	Recommended local starting point	Local model memory unnecessary
CPU only · 32–64 GB RAM	Possible, slower bot turns	Local GPU unnecessary
Rented GPU · 48 GB VRAM	Not used for local route	Initial vLLM memory budget; verify worker startup

The 15.7 GB file size is not total runtime memory. Do not apply llama.cpp memory measurements directly to vLLM. Additional context, concurrent bots and Docker need headroom.

SELF-HOSTED BOTS / LOCAL OR RENTED INFERENCESOURCE-LINKED SETUP

Create Uncensored AI Bots with Rakazo

Self-host Rakazo, connect OrcaSAQ-2-Cyber-27B-Uncensored-GGUF, and test a real bot action. Choose local Ollama or RunPod Serverless with vLLM. Rakazo stays on your computer in both routes. The Serverless route needs a GGUF-enabled worker; the download below supplies that extra setup.

Choose a package Run its commands Check the result

BEFORE YOU BEGIN

The documented package

This setup uses the package documented for this task. Review its source and supported platforms before starting.

These requirements describe the GPU you rent. Use Windows, Mac or Linux as the browser client; your saved computer’s VRAM does not limit this setup.

CURRENT SETUP

OrcaSAQ-2 Cyber 27B · uncensored · community-modified weights

Local Ollama: start at 8192 context, 24 GB GPU or 32 GB unified memory. Serverless vLLM: budget 48 GB VRAM initially; GGUF/runtime overhead must be checked on the worker. The model’s advertised maximum context is not a hardware fit guarantee.

Freshness unavailable. Date pending.

Popularity unavailable. Stats pending.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 17, 2026. Statistics check pending. Statistics are marked for refresh after 72 hours. These statistics are due for a refresh.Full scoring method

View weight source

THE WORKFLOW

Set up OrcaSAQ-2 Cyber 27B · uncensored from your own computer

Run the app on your own computer; only model inference runs on the rented GPU. The instructions below update to match your selection.

CHOOSE YOUR SETUPRented GPU

RunPod Serverless + vLLM, using the exact OrcaSAQ GGUF. Rakazo remains on your own computer. The supplied worker adds vLLM’s required GGUF plugin; it has not been GPU-tested end to end.

Windows: enable WSL2 and Docker Desktop

Windows users only; Mac/Linux users skip to the next step. If Ubuntu in WSL2 is not installed, run the command below in PowerShell as Administrator, restart if requested, then open Ubuntu and create its Linux username/password. Install Docker Desktop, choose its WSL2 engine, and enable Settings → Resources → WSL Integration → Ubuntu. From here, run Rakazo’s Bash commands in Ubuntu, not PowerShell. Keep the installation under ~/rakazo in Ubuntu.

Install Docker Desktop for Windows Docker’s WSL2 integration instructions

PowerShell

wsl --install -d Ubuntu

Mac or Linux: start Docker; Windows: continue in Ubuntu

Mac: install and start Docker Desktop, then open Terminal. Linux: install Docker Engine 26+ with Compose and allow your user to use Docker. Windows: use the Ubuntu shell from the previous step with Docker Desktop running. On Ubuntu/WSL, install missing prerequisites with the command below; macOS needs curl and openssl but does not use apt-get.

Docker Desktop for Mac Docker Engine for Linux

Bash

sudo apt-get update
sudo apt-get install -y curl openssl

Install Rakazo once

Run this in macOS/Linux Terminal or Windows Ubuntu/WSL2. The official installer downloads published Docker images, creates .env with random secrets, and starts the app. It uses the edge image tag by default; record your image version for a repeatable demo. No source build is needed.

Official Rakazo published-image installation

Bash

mkdir -p ~/rakazo && cd ~/rakazo
curl -fsSLO https://raw.githubusercontent.com/elie222/rakazo/main/infra/compose/install-images.sh
bash install-images.sh

Open Rakazo and create your account

Open http://127.0.0.1:5173 in your computer’s browser. The first registered account becomes the deployment owner. Docker bot computers work by default; no E2B, Daytona, or Box key is required. Keep the default local-only address for this first setup.

Prepare the GGUF-enabled Serverless worker once

This exact model is a GGUF, so the stock vLLM worker needs the GGUF plugin and a download step. Download and extract the small worker bundle below. Install Git and Docker Desktop, then open Terminal or Windows Ubuntu/WSL2 in the extracted folder. Run the commands below to build the pinned official RunPod worker and add GGUF support. Docker must be running. This builds an image; it does not rent a GPU. The current vLLM documentation marks GGUF experimental, so keep the startup check below.

Download OrcaSAQ Serverless worker files Install Git vLLM GGUF requirements

Bash

git clone https://github.com/runpod-workers/worker-vllm.git runpod-vllm
git -C runpod-vllm checkout ab52ac322daf1d3c3111f62af7daa26c1d4d715e
docker build --platform linux/amd64 -t rakazo-runpod-vllm-base:local ./runpod-vllm
docker build --platform linux/amd64 -t YOUR_DOCKERHUB_USER/rakazo-orca-vllm:1 .

Push the worker image

Create a Docker Hub account/repository if needed. Replace YOUR_DOCKERHUB_USER in both build and push commands with your username. Sign in using Docker’s login flow, then push the image. Never put Hugging Face or RunPod keys in the image. Use this image in RunPod Serverless—not in Pods.

Docker Hub

Bash

docker login
docker push YOUR_DOCKERHUB_USER/rakazo-orca-vllm:1

Create the RunPod Serverless endpoint

Open RunPod → Serverless → New Endpoint and choose your custom container image. Use YOUR_DOCKERHUB_USER/rakazo-orca-vllm:1. Start with one 48 GB GPU worker, maximum workers 1 and active/minimum workers 0. That limits concurrency and allows scale-to-zero. Add the environment settings in the next step before starting. No exposed Pod ports, cluster, SSH tunnel or llama.cpp server is involved.

Open RunPod · affiliate link RunPod Serverless custom workers

Set the OrcaSAQ model and tool settings

Add these values under endpoint environment settings. HF_TOKEN must be an authorized read token after accepting the exact model repository’s access conditions. The startup script downloads only OrcaSAQ-2-27B-Uncensored.gguf and sets MODEL_NAME to that file. Qwen3.8 below supplies tokenizer/config metadata for OrcaSAQ’s base architecture—it does not select different model weights. Keep the token private. These are configuration starting points, not a tested compatibility guarantee.

Exact OrcaSAQ model and access conditions Official RunPod worker configuration

Plain text

HF_TOKEN=YOUR_AUTHORIZED_HUGGING_FACE_READ_TOKEN
TOKENIZER=orcarouter/Qwen3.8-27B-Uncensored
HF_CONFIG_PATH=orcarouter/Qwen3.8-27B-Uncensored
OPENAI_SERVED_MODEL_NAME_OVERRIDE=orcasaq-cyber-27b
MAX_MODEL_LEN=8192
MAX_NUM_SEQS=1
ENABLE_AUTO_TOOL_CHOICE=true
TOOL_CALL_PARSER=qwen3_coder
REASONING_PARSER=qwen3

Wait for the worker to load OrcaSAQ

Create the endpoint, copy its Endpoint ID, and obtain a RunPod API key. Inspect worker logs until vLLM finishes loading the GGUF and becomes ready. If startup reports an unsupported tensor/architecture, plugin error or missing config, do not repeatedly retry a paid worker: pause the endpoint and retain the exact error. This GGUF’s compatibility with the worker must be established by this startup check; loading successfully is still separate from the bot tool test.

Allow the public model API in Rakazo

In your existing ~/rakazo folder, open .env in a text editor and add or update exactly one line as shown. Preserve all existing secrets. This allows a public HTTPS OpenAI-compatible model endpoint; it does not publish Rakazo itself. Keep .env and API keys out of source control and video recordings.

Rakazo OpenAI-compatible endpoint settings

Plain text

RAKAZO_OPENAI_COMPAT_ALLOW_PUBLIC=1

Apply the setting to the existing stack

From the same Rakazo directory, recreate services using the existing Compose file. This keeps the configured image and volumes; no reinstall is needed.

Bash

cd ~/rakazo
docker compose --env-file .env -f docker-compose.images.yml up -d

Connect the OrcaSAQ Serverless endpoint in Rakazo

Go to Settings → Models → Connect a model → OpenAI-compatible. Use the Endpoint ID from RunPod and your RunPod API key—not HF_TOKEN. Save the connection, create a bot and choose OrcaSAQ. Leave Supports images off. Use /openai/v1, not /run or /runsync.

Setting	Value
Base URL	https://api.runpod.ai/v2/YOUR_ENDPOINT_ID/openai/v1
Model ID	orcasaq-cyber-27b
API key	Your RunPod API key
Supports images	Off

Create a bot and test a real tool action

Create a bot and select the model connection you just saved. Keep its computer provider on Docker and use a test workspace without sensitive files or credentials. Send the prompt below. Open Agent computer / its file view and inspect test.txt and the tool activity. A reply claiming it created a file is not enough: the file must exist, and its contents must match the read-back. OrcaSAQ-2 Cyber 27B is text-only: use shell/file tools, not screenshot-based browser tasks.

Plain text

Create a file named test.txt containing today's date, then read that file and report exactly what it contains. Use your file or shell tools; do not just describe the actions.

MORE WAYS TO USE THIS SETUP

Stop, restart, and keep your bots

Use the same ~/rakazo directory and .env on every restart. Windows users run these commands in Ubuntu/WSL2; Mac/Linux users use Terminal.

Stop the local stack

This stops Rakazo’s Compose services without deleting its data volumes. Stop any active bot work first. It does not disable a RunPod endpoint: manage remote workers separately in RunPod to prevent unwanted charges.

Bash

cd ~/rakazo
docker compose --env-file .env -f docker-compose.images.yml stop

Start again without reinstalling

Start Docker first, then run this from your existing folder. For local inference, also start Ollama. Open http://127.0.0.1:5173 and sign in to the same account. For RunPod, check that its endpoint is still available; the first request may have a cold start.

Bash

cd ~/rakazo
docker compose --env-file .env -f docker-compose.images.yml up -d

Preserve the data that matters

Keep .env and the Docker pgdata and appdata volumes. They hold the database and application/computer data; stop/start preserves volumes. Do not use docker compose down -v or delete these volumes. Keep backups of the volumes and the original encryption secrets together. The installer uses edge images by default; rerunning it can pull an update, so use the restart command above for routine starts.

MORE WAYS TO USE THIS SETUP

Use the same bots from your phone

The phone connects to Rakazo—not directly to your model endpoint. RunPod in this guide hosts inference only.

Connect to the same backend

On the mobile sign-in screen choose Use a custom server and enter your reachable Rakazo server origin. The default 127.0.0.1 address is accessible only on the host, not from your phone. For access away from home, follow Rakazo’s HTTPS self-hosting setup and sign in to the same account; do not simply expose the local port. Keep the host awake, or move the backend to an always-on server.

Rakazo mobile releases HTTPS hosting and mobile connection instructions

MORE WAYS TO USE THIS SETUP

Optional connection checks

Use these only when the connection fails. They are diagnostics, not extra installation steps.

Check local Ollama

On Mac use Terminal; on Windows use normal PowerShell (curl.exe instead of curl). On Linux replace the URL with the private listener address configured in the Linux steps. The response should list the model.

Bash

curl http://127.0.0.1:11434/v1/models

Check the RunPod endpoint

In macOS/Linux Terminal or Ubuntu/WSL2, enter the Endpoint ID and API key when prompted. Key input is hidden. A successful response lists orcasaq-cyber-27b; use that exact ID in Rakazo. Never record or share your key.

Bash

bash
read -r -p "RunPod endpoint ID: " ENDPOINT_ID
read -r -s -p "RunPod API key: " RUNPOD_API_KEY; printf "\n"
curl -sS "https://api.runpod.ai/v2/${ENDPOINT_ID}/openai/v1/models" -H "Authorization: Bearer ${RUNPOD_API_KEY}"
unset RUNPOD_API_KEY

Inspect Rakazo services

Check service status and recent startup errors. Keep log output private if it contains credentials or prompts.

Bash

cd ~/rakazo
docker compose --env-file .env -f docker-compose.images.yml ps
docker compose --env-file .env -f docker-compose.images.yml logs --tail=80 api worker

SUCCESS CHECK · OrcaSAQ-2 Cyber 27B · uncensored

A real file action—not just a chat reply

Rakazo connects to the chosen model, the model emits a usable tool call, and the bot computer creates test.txt. Inspect the actual file, then compare its contents with the bot’s read-back. All three must work before calling the integration successful.

http://127.0.0.1:5173Opens on your machine

The local and Serverless routes use the exact OrcaSAQ GGUF. Documentation and page checks are not a live GPU test. vLLM GGUF support is experimental; this exact worker/model integration has not been run end to end by YouRunAI. Confirm startup and the file-action test before recording a successful demo.

Back to setup steps

When it doesn’t go to plan

The local page does not open

Confirm Docker is running and check the service status in Optional connection checks. Use the same ~/rakazo folder; inspect startup logs before reinstalling. If port 5173 is occupied, set RAKAZO_WEB_PORT to a free port in .env and recreate services, then open that port.

Ollama responds in Terminal but not in Rakazo

Docker Desktop: use host.docker.internal:11434/v1, not localhost. Native Linux: follow the private-IP listener steps and verify container-to-host firewall access. Do not expose an unauthenticated model port publicly.

Rakazo rejects the RunPod address

Use HTTPS and set RAKAZO_OPENAI_COMPAT_ALLOW_PUBLIC=1 once in the existing .env. Apply it with the Compose up -d command so both API and worker receive the change.

The RunPod worker cannot download weights

Open the exact Hugging Face repository, accept its access conditions if you agree, and give the worker an HF_TOKEN secret authorized for it. A RunPod API key cannot unlock Hugging Face files.

RunPod returns 401 or 403 to Rakazo

Use the RunPod API key in Rakazo, check the Endpoint ID, and confirm the URL ends in /openai/v1. Do not use the Hugging Face download token as API authorization.

The worker loads but replies are delayed or fail

Check cold-start/worker logs and available GPU memory. Start at 8192 context with one bot. Reduce context or rent more memory if the worker reports out of memory; avoid multiple simultaneous test requests.

The model chats but does not create files

Check that the Docker bot computer is available and inspect actual tool calls. For vLLM, confirm ENABLE_AUTO_TOOL_CHOICE=true, TOOL_CALL_PARSER=qwen3_coder and REASONING_PARSER=qwen3. OrcaSAQ may still fail tool use: parsing support alone is not proof of agent reliability. Use text/file tasks with Supports images off.

A longer conversation fails or forgets instructions

Model context includes the system prompt, tool definitions, history and response. Begin with a fresh short task. Increase Ollama’s context allocation only with spare memory; RunPod’s MAX_MODEL_LEN=8192 is a starting cap, not an unlimited history. See the linked context documentation.

The phone has different history or cannot connect

Use the same reachable Rakazo backend and account. localhost on your phone is not your computer. The host must remain awake and reachable; a RunPod model endpoint does not synchronize Rakazo history.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

9 SOURCES

01Rakazo · repository and quick startgithub.com 02Rakazo · self-hosting and connectionsgithub.com 03OrcaSAQ · exact GGUF modelhuggingface.co 04Ollama · GGUF importdocs.ollama.com 05Ollama · downloadsollama.com 06RunPod · Serverless vLLMdocs.runpod.io 07RunPod · OpenAI-compatible endpointdocs.runpod.io 08vLLM · GGUF plugin requirementsdocs.vllm.ai 09OrcaSAQ · downloadable Serverless workeryourunai.com

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat setup

Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.

Freshness 41 out of 100. Release age: 232 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 75 out of 100. Repository age: 73 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 10 of 57. Ranked model: Qwen3.8-27B. Original model ranking, not a benchmark of this package. Export dated Oct 16, 2026. Source checked Oct 17, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Create images locally with FLUX.2 Klein

Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.

Freshness 35 out of 100. Release age: 275 days old, Jan 15, 2026. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source